Boris Cherny Files
The exact subagents and slash commands Claude Code's creator runs on every PR. Two Anthropic-official, five community-written.
What's in this pack
This pack collects the seven Claude Code assets that match the personal workflow Boris Cherny — the creator and tech lead of Claude Code at Anthropic — describes on his public setup page (howborisusesclaudecode.com) and in his Threads posts. Two are official Anthropic releases. Five are community-written equivalents of the slash commands Boris demos in his talks.
| # | Asset | Type | Source |
|---|---|---|---|
| 1 | Ralph Wiggum | slash command | Anthropic official plugin |
| 2 | code-simplifier | subagent | Anthropic official subagent |
| 3 | verify-app | subagent | community equivalent |
| 4 | code-architect | subagent | community equivalent |
| 5 | sentry-errors | subagent | community equivalent |
| 6 | oncall-guide | subagent | community equivalent |
| 7 | build-validator | subagent | community equivalent |
The two Anthropic-official entries are the same files Boris uses. The five community subagents reproduce the behavior of the workflows he demonstrates publicly — they are not copies of his private files.
Why this matters
Watching Boris's setup walkthrough is one thing. Reproducing it across your projects is another. Most engineers see the demo, write down the names, and never get around to wiring them up. This pack closes that gap: you install once, and every Claude Code session you launch picks up the same .claude/agents/*.md files and .claude/commands/*.md files that mirror Boris's daily flow.
The two official assets are the load-bearing ones:
- Ralph Wiggum turns Claude Code into an autonomous loop. You give it a prompt and a completion sentinel, and it iterates — running tests, fixing failures, re-running — until it emits the sentinel or hits the iteration cap. Boris uses it for his longest-running tasks (multi-file refactors, "until-it-passes" TDD).
- code-simplifier is the cleanup pass Boris runs after almost every Claude Code session. It refactors recently modified files for clarity without changing observable behavior. Anthropic open-sourced it in late 2025 from the internal team's own setup.
The five community subagents fill in the rest of the loop:
- verify-app runs end-to-end smoke tests after a change.
- code-architect does PR-grade structural reviews before commit.
- sentry-errors fetches recent production errors and prioritizes a fix list.
- oncall-guide generates incident runbooks from a description.
- build-validator checks builds before push.
Install in one command
# Install the entire pack into the current Claude Code project
tokrepo install pack/boris-cherny-files
# Or pick individual assets
tokrepo install ralph-wiggum
tokrepo install code-simplifier
The TokRepo CLI normalizes file placement across all eight supported AI tools (Claude Code, Cursor, Codex CLI, Gemini CLI, Cline, Roo Code, Windsurf, GitHub Copilot). Install once, and Claude Code picks them up on the next session.
When to use which
| Situation | Asset |
|---|---|
| Long-running multi-file refactor | Ralph Wiggum (set --max-iterations) |
| End of any session where Claude wrote >100 lines | code-simplifier |
| Pre-PR structural review | code-architect |
| Production incident triage | sentry-errors → oncall-guide |
| Pre-push gate | build-validator |
| End-to-end smoke after a feature | verify-app |
Cost discipline
Ralph Wiggum can burn through a token budget if mis-configured. Always set --max-iterations and make --completion-promise something Claude can only emit when verification passes (e.g. all tests green). Start with small tasks; observe the iteration count before scaling up.
7 assets in this pack
Frequently asked questions
Are these the actual files Boris Cherny uses internally?
Two of them — Ralph Wiggum and code-simplifier — are the actual Anthropic-official releases that Boris ships with. The other five are community-written equivalents that reproduce the behavior of the slash commands he demos publicly, not copies of his private files. We document this distinction on every asset page so you never misattribute.
Will this work with Cursor, Codex CLI, or Windsurf instead of Claude Code?
The TokRepo CLI normalizes file placement across all 8 supported AI tools. Subagents map to the closest equivalent (Cursor rules for Cursor, AGENTS.md for Codex CLI, etc). Some assets are Claude-Code-specific by design — Ralph Wiggum is a Claude Code plugin and won't run elsewhere. Each asset's page lists the tools it supports.
How does the one-command install actually work?
tokrepo install pack/boris-cherny-files reads the pack manifest, fetches each asset's prompt template, and writes them to .claude/agents/ and .claude/commands/ in your current project. No global install, no daemon — just files dropped into the right place. Run tokrepo uninstall
What's the difference between Ralph Wiggum and a regular slash command?
A slash command runs once and returns. Ralph runs in a loop, re-feeding Claude Code its own prompt + completion criteria each iteration, until either the model emits a completion sentinel string or --max-iterations is hit. It's designed for tasks where you'd otherwise babysit Claude over many turns (long refactors, TDD).
Will code-simplifier change my code's behavior?
No. The agent's contract is to never change observable behavior — only structure. Public APIs, return values, side effects, and edge-case handling all stay intact. It commits cleanups as a separate diff so you can review before merging.
12 packs · 80+ hand-picked assets
Browse every curated bundle on the home page
Back to all packs